Why the Next AI Winner Will Not Come From Silicon Valley. It Will Come From Here

There is a version of the AI story that is frequently recounted: a group of San Francisco-based founders wearing hoodies, a few billion dollars from Sequoia, an internet-breaking product launch, and an incomprehensible valuation. It’s an interesting tale. It’s also beginning to feel lacking.

It’s possible that the next big AI company won’t be located within twenty miles of Caltrain. It might originate from a Detroit manufacturing floor where a fifty-year-old engineer has been silently gathering the kind of specific, messy, irreplaceable data that no language model trained on the open internet will ever have. The engineer has been staring at the same broken asset register for ten years. Alternatively, it might originate from London, where Demis Hassabis created Google DeepMind from the ground up while Silicon Valley investors passed him by in search of the next San Jose office park to finance. Hassabis recently stated, “I knew there was talent here being overlooked in the U.S.,” at a London event. He was correct. It’s also noteworthy that it took the Valley ten years to fully recognize it.

Because Hassabis’s path was so intentional, it is instructive. When he started DeepMind in 2010, he was already certain that artificial intelligence would develop into what it is today. This belief required a certain amount of perseverance to endure years when, in his words, “nothing was working.” Relocation was the obvious choice when Google bought the business in 2014. Mountain View, or at least a peninsula location. Hassabis remained in London. He recently opened a new office there, a spacious, luxurious location just a short distance from the original, because, as he has stated time and time again, the people who make AI decisions shouldn’t all reside in the same twenty square miles of California. It’s a valid point that receives less attention than it merits.

However, geography isn’t really at the heart of the deeper argument. It concerns the type of AI that genuinely generates long-term value. Building foundation models was the focus of the first wave, which continues to dominate headlines. Large, costly, all-purpose systems that are versatile and trained on everything. Amazing things came out of that phase. Additionally, it created a situation where the businesses that spend the most on infrastructure are still having trouble articulating the true return on that investment. For a large portion of the first part of 2026, investors asked that question in private. Some people continue to inquire about it.

Alongside that uncertainty, another kind of opportunity is emerging. Not model-building, but application-building—that is, using AI to solve the genuinely challenging, unglamorous issues present in sectors that Silicon Valley has historically deemed dull. supply chains with data that is inconsistent over decades. Records in hospital systems were never intended to communicate with one another. manufacturing facilities where engineers with thirty years of experience hold the most valuable institutional knowledge. A general-purpose chatbot cannot solve these issues. These are issues that call for patience, domain knowledge, and a willingness to spend six months cleaning data before the AI produces any interesting results. That type of work has historically caused allergies in the Valley. Other locations haven’t been.

The most obvious American example is Michigan. Earlier this month, Crain’s Detroit published an article arguing that the state’s advantage in AI comes from people who understand “the messiness of the real world”—manufacturing engineers, supply chain managers, and field service technicians who have spent careers bridging the gap between how systems are supposed to work and how they actually do—rather than from compute or venture capital. Framing could come across as defensive, like a local publication promoting its own interests. Maybe it’s just right.

China, on the other hand, has been quietly moving in a different direction, concentrating on practical deployment and open-source, energy-efficient models rather than chasing the raw parameter counts that American labs have made their standard for advancement. It’s still unclear if that strategy results in a truly competitive general AI or just a very useful collection of industrial tools. Both results would be important.
As all of this is happening, it’s difficult to ignore how the AI story is breaking up in ways that don’t neatly fit into the narrative of one nation, one city, or one group of businesses winning everything. It is carefully framed by Hassabis, who is not prone to false modesty: the people creating AI shouldn’t only come from twenty square miles of the United States. The world will be impacted by the technology. The locations constructing it most likely ought to as well. In five years, that argument will appear either clear-cut or revolutionary. It currently sits awkwardly in the middle, which is typically where the interesting things are.